Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis